Inspiration
The same spreadsheet mistakes return every week. A comparison can show what changed, but it cannot by itself preserve what a person meant by a correction. Mendline keeps that lesson while leaving each later decision with the user.
What it does
Teach a small correction, clarify the intended rule, inspect its scope, then preview and explicitly approve its application. On a later export, the lesson returns as an inactive proposal requiring fresh approval. Uncertain record identities and competing corrections are held for review. Original files stay unchanged.
Try it
Open https://mendline.pages.dev and choose Try the example yourself. Correct SKU:tea to tea, answer what SKU:SKU:tea should become, and choose removal of one leading SKU: with an explicit starts-with scope. A genuinely later export has reordered columns and the unseen value SKU:novel. Review its proposed change to novel, keep ambiguous and competing records unchanged, then download the corrected CSV and separate history. No account or API key is needed.
How we built it
TypeScript, React and Vite provide a static browser-local workbench. A small allowlisted rule grammar produces alternatives; no arbitrary code or runtime generative AI executes. An immutable review lifecycle connects original, corrected and latest snapshots with identity guards, explicit approvals, preview, undo, recovery and export.
Kiro use
Kiro IDE performed native source reading and specification writing on 2 November 2026. The genuine requirements and design are included at .kiro/specs/correction-continuity/. They were reviewed and corrected before later implementation, including excluding key transformations. These historical specifications are not a claim that Kiro wrote all implementation or ran tests. Additional engineering assistants implemented and reviewed the application.
Challenges and lessons
One example can fit several operations: stripping a prefix and replacing a literal may both explain the same correction. We learned to expose uncertainty, ask a distinguishing question, default to narrow scope and require a separate application decision. Reordered exports also need stable identity assumptions; equal-looking IDs alone are not enough. Recovery and undo must never silently replay old approval onto replacement files.
What we verified
The release has 143 passing focused tests plus typechecking and build checks. Actual browser acceptance includes rule clarification, fresh later-file approval, held conflicts and downloaded output. The public deployment completed the seven-stage example; its saved CSV matches the accepted expected result byte-for-byte. The captioned demo is an edited recording of real browser interaction, not a simulated result.
New work and prior work
Mendline is a separate project whose local history begins 2 November 2026. Earlier CSV Change Review work informed the problem: keyed CSV comparison, cell evidence, human notes and proposal staging are prior capabilities, not new inventions. Mendline adds the O/C/N correction overlay, finite rule teaching and clarification, inactive later-export reuse with fresh approval, and corrected CSV export. This scoped comparison is not a claim of global novelty. Original code is MIT licensed; dependency notices are included.
Limits and next steps
This is bounded correction intelligence, not general-purpose learning. Identity continuity remains a user-declared assumption; no fuzzy matching is claimed. Histories and recovery may contain raw values and are not encrypted. Use fictional or otherwise permitted nonconfidential data. Next we would measure whether first-time users can complete the correction workflow reliably before expanding its rule language.
Log in or sign up for Devpost to join the conversation.